| Message ID | 014a655b-7080-3804-3a56-5e00851a2a7d@suse.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | x86: build adjustments | expand |
On 07/08/2020 12:33, Jan Beulich wrote: > While this won't affect overall memory overhead (struct vcpu as well as > struct domain get allocated as single pages) nor code size (the offsets > into the base structures are too large to be representable as signed 8- > bit displacements), it'll allow the tail of struct pv_{domain,vcpu} to > share a cache line with subsequent struct arch_{domain,vcpu} fields. > > Signed-off-by: Jan Beulich <jbeulich@suse.com> > --- > RFC: There is a risk associated with this: If we still have code > somewhere accessing the HVM parts of the structures without a prior > type check of the guest, this going to end up worse than the so far > not uncommon case of the access simply going to space unused by PV. > We may therefore want to consider whether to further restrict when > this conversion to union gets done. > And of course there's also the risk of future compilers complaining > about this abuse of unions. But this is limited to code that's dead > in !HVM configs, so only an apparent problem. > > --- a/xen/include/asm-x86/hvm/domain.h > +++ b/xen/include/asm-x86/hvm/domain.h > @@ -99,7 +99,13 @@ struct hvm_pi_ops { > > #define MAX_NR_IOREQ_SERVERS 8 > > -struct hvm_domain { > +typedef > +#ifdef CONFIG_HVM > +struct > +#else > +union > +#endif > +hvm_domain { > /* Guest page range used for non-default ioreq servers */ > struct { > unsigned long base; > @@ -203,7 +209,7 @@ struct hvm_domain { > #ifdef CONFIG_MEM_SHARING > struct mem_sharing_domain mem_sharing; > #endif > -}; > +} hvm_domain_t; Honestly, I'd say no to this patch even it resulted in a 100x speedup, because I am totally lost for words about this construct, and the effect it comprehensibility of our code. Seeing as improved cache locality appears to be the sole benefit, marginal as it is, you can achieve that in a better way by rearranging struct domain/vcpu to put the pv/hvm union at the end. ~Andrew
On 07.08.2020 19:14, Andrew Cooper wrote: > On 07/08/2020 12:33, Jan Beulich wrote: >> While this won't affect overall memory overhead (struct vcpu as well as >> struct domain get allocated as single pages) nor code size (the offsets >> into the base structures are too large to be representable as signed 8- >> bit displacements), it'll allow the tail of struct pv_{domain,vcpu} to >> share a cache line with subsequent struct arch_{domain,vcpu} fields. >> >> Signed-off-by: Jan Beulich <jbeulich@suse.com> >> --- >> RFC: There is a risk associated with this: If we still have code >> somewhere accessing the HVM parts of the structures without a prior >> type check of the guest, this going to end up worse than the so far >> not uncommon case of the access simply going to space unused by PV. >> We may therefore want to consider whether to further restrict when >> this conversion to union gets done. >> And of course there's also the risk of future compilers complaining >> about this abuse of unions. But this is limited to code that's dead >> in !HVM configs, so only an apparent problem. >> >> --- a/xen/include/asm-x86/hvm/domain.h >> +++ b/xen/include/asm-x86/hvm/domain.h >> @@ -99,7 +99,13 @@ struct hvm_pi_ops { >> >> #define MAX_NR_IOREQ_SERVERS 8 >> >> -struct hvm_domain { >> +typedef >> +#ifdef CONFIG_HVM >> +struct >> +#else >> +union >> +#endif >> +hvm_domain { >> /* Guest page range used for non-default ioreq servers */ >> struct { >> unsigned long base; >> @@ -203,7 +209,7 @@ struct hvm_domain { >> #ifdef CONFIG_MEM_SHARING >> struct mem_sharing_domain mem_sharing; >> #endif >> -}; >> +} hvm_domain_t; > > Honestly, I'd say no to this patch even it resulted in a 100x speedup, > because I am totally lost for words about this construct, and the effect > it comprehensibility of our code. Okay - away it goes. Jan
--- a/xen/arch/x86/domctl.c +++ b/xen/arch/x86/domctl.c @@ -709,7 +709,7 @@ long arch_do_domctl( unsigned int fmp = domctl->u.ioport_mapping.first_mport; unsigned int np = domctl->u.ioport_mapping.nr_ports; unsigned int add = domctl->u.ioport_mapping.add_mapping; - struct hvm_domain *hvm; + hvm_domain_t *hvm; struct g2m_ioport *g2m_ioport; int found = 0; --- a/xen/include/asm-x86/domain.h +++ b/xen/include/asm-x86/domain.h @@ -319,7 +319,7 @@ struct arch_domain union { struct pv_domain pv; - struct hvm_domain hvm; + hvm_domain_t hvm; }; struct paging_domain paging; @@ -582,7 +582,7 @@ struct arch_vcpu /* Virtual Machine Extensions */ union { struct pv_vcpu pv; - struct hvm_vcpu hvm; + hvm_vcpu_t hvm; }; /* --- a/xen/include/asm-x86/hvm/domain.h +++ b/xen/include/asm-x86/hvm/domain.h @@ -99,7 +99,13 @@ struct hvm_pi_ops { #define MAX_NR_IOREQ_SERVERS 8 -struct hvm_domain { +typedef +#ifdef CONFIG_HVM +struct +#else +union +#endif +hvm_domain { /* Guest page range used for non-default ioreq servers */ struct { unsigned long base; @@ -203,7 +209,7 @@ struct hvm_domain { #ifdef CONFIG_MEM_SHARING struct mem_sharing_domain mem_sharing; #endif -}; +} hvm_domain_t; #endif /* __ASM_X86_HVM_DOMAIN_H__ */ --- a/xen/include/asm-x86/hvm/vcpu.h +++ b/xen/include/asm-x86/hvm/vcpu.h @@ -149,7 +149,13 @@ struct altp2mvcpu { #define vcpu_altp2m(v) ((v)->arch.hvm.avcpu) -struct hvm_vcpu { +typedef +#ifdef CONFIG_HVM +struct +#else +union +#endif +hvm_vcpu { /* Guest control-register and EFER values, just as the guest sees them. */ unsigned long guest_cr[5]; unsigned long guest_efer; @@ -213,7 +219,7 @@ struct hvm_vcpu { struct x86_event inject_event; struct viridian_vcpu *viridian; -}; +} hvm_vcpu_t; #endif /* __ASM_X86_HVM_VCPU_H__ */
While this won't affect overall memory overhead (struct vcpu as well as struct domain get allocated as single pages) nor code size (the offsets into the base structures are too large to be representable as signed 8- bit displacements), it'll allow the tail of struct pv_{domain,vcpu} to share a cache line with subsequent struct arch_{domain,vcpu} fields. Signed-off-by: Jan Beulich <jbeulich@suse.com> --- RFC: There is a risk associated with this: If we still have code somewhere accessing the HVM parts of the structures without a prior type check of the guest, this going to end up worse than the so far not uncommon case of the access simply going to space unused by PV. We may therefore want to consider whether to further restrict when this conversion to union gets done. And of course there's also the risk of future compilers complaining about this abuse of unions. But this is limited to code that's dead in !HVM configs, so only an apparent problem.